CN-121997762-A - Forging heat coordinated material organization and performance integrated intelligent regulation and control method and system

CN121997762ACN 121997762 ACN121997762 ACN 121997762ACN-121997762-A

Abstract

The invention discloses a forging heat coordinated material organization and performance integrated intelligent regulation and control method and system, and belongs to the technical field of material processing and intelligent manufacturing. The method comprises the steps of 1) carrying out parameter calibration on a multi-module coupled physical model by using a small amount of real experimental data, 2) generating a large amount of virtual process combination parameter data in a process parameter space by using a Latin hypercube sampling strategy, inputting the parameter data into the calibrated physical model for calculation to obtain an output result, screening the output result by taking physical constraint as a condition to obtain an enhanced data set meeting the physical constraint, and 3) training and cascading the predicted physical model by using the enhanced data set. The hierarchical prediction physical model realizes the integrated and high-precision prediction of a plurality of macroscopic properties and microscopic structure indexes of the material, ensures the physical consistency of the prediction result, ensures the prediction capability of the physical derivative characteristic contribution exceeding 70 percent, and ensures the interpretability of the model.

Inventors

CHENG PENG
ZHOU JUNYING
SHAO CHENXI
CHENG LU
Guo Yehang

Assignees

中机生产力促进中心有限公司

Dates

Publication Date: 20260508
Application Date: 20260129

Claims (10)

1. The forging and heating synergistic material organization and performance integrated intelligent regulation and control method is characterized by comprising the following steps of: 1) Parameter calibration is carried out on the physical model coupled by the multiple modules by using a small amount of real experimental data; 2) Generating a large amount of virtual process combination parameter data in a process parameter space by using a Latin hypercube sampling strategy, inputting the parameter data into a calibrated physical model for calculation to obtain an output result, and screening the output result by taking physical constraint as a condition to obtain an enhanced data set meeting the physical constraint; 3) And training a cascade predictive physical model by using the enhanced data set.
2. The method according to claim 1, wherein in the step 3), the cascade predictive physical model is a three-stage cascade predictive architecture established according to a physical causal chain of carbon content- & gt hardness- & gt depth of percolation layer; The training specifically comprises the following three stages: the first stage, surface carbon content prediction, is to adopt four model integrated learning (gradient lifting regression, random forest, extreme random tree, ridge regression), input as a combined vector of technological parameters and physical derivative characteristics, and output as a surface carbon content predicted value ; The second stage, maximum hardness prediction, adopts a physical baseline and residual error learning architecture. The physical baseline is provided by the carbon-hardness equation: The machine learning model learns residual correction terms H ˆ v=hv physics +∆HV ML ; And the third stage, effective seepage depth prediction, adopts a 'calibration Fick law+residual error learning' framework. The physical baseline is provided by the calibrated depth model, x physics = α 2 ∈dt-erfc-1 (Ccrit/Cs), the final predicted value: ; the training cascade predictive physical model uses an enhanced data set as a training set to define a joint loss function: Wherein L MSE is a mean square error loss, used for regression tasks, L CE is a cross entropy loss, used for classification tasks, and w i and lambda j are weight coefficients of the tasks; The training cascade prediction physical model only uses regression parts in the loss function, i=1, 2 and 3 respectively correspond to three regression targets of surface carbon content, maximum hardness and effective percolation layer depth, and the weight coefficient is set to w1=1.0, w2=0.01 and w3=1.0.
3. The method according to claim 2, characterized in that: In the step 3), training a classification target prediction physical model architecture by utilizing the enhanced data set, specifically, adopting a multi-task learning architecture, sharing characteristics to extract a backbone network, and differentiating a plurality of classification pre-measuring heads at the tail end for five classification targets of an internal oxidation level, a martensite level, a retained austenite level, a carbide level, a core tissue level and the like; Specifically, only the classification part in the loss function is used, wherein j=1, 2,..5 corresponds to 5 classification targets respectively, and the weight coefficient is set according to the difficulty and importance of each classification task, and is lambda_j=1.0 by default.
4. The method according to claim 1, characterized in that: the physical model in step 1) includes: and (3) calculating a concentration distribution curve of carbon on the surface layer of the steel based on Fick second law and an error function solution, wherein the concentration distribution curve of carbon on the surface layer of the steel under different process parameters is expressed as: Wherein the diffusion coefficient D (T) follows the Arrhenius equation: Carbon-hardness conversion model based on martensite solid solution strengthening mechanism, establishing the relation between carbon content and Vickers hardness by linear regression: HV=H 0 + H C ·C(5) Wherein H 0 is matrix hardness, and H C is hardening coefficient of carbon; Effective depth of layer model based on Fick's law, calculate effective depth of hardened layer: where α is the calibration factor and Ccrit is the critical carbon content.
5. The method according to claim 1, characterized in that: the physical constraints include: carbon content constraint of 0.3wt% < C <1.2wt%; Hardness constraint 500HV <900HV; The depth of seepage constraint is 0.5mm < x <2.0mm.
6. The method according to claim 1, characterized in that: and constructing physical derivative characteristic data on the basis of the enhancement data, wherein the physical derivative characteristic data comprises a characteristic diffusion length V & ltDt, a logarithmic diffusion coefficient log10 (D) and a diffusion ratio tboost/tdiff.
7. The method according to claim 1, characterized in that: and 4) model verification, wherein the cross verification is carried out by adopting a two-stage verification method of 5-fold cross verification and leave-one-out cross verification of the enhanced data.
8. The method according to claim 1, characterized in that: Also included is step 5) model tuning to adapt to new materials, comprising the sub-steps of: recalibrating parameters of a physical model, carrying out parameter calibration by using a small amount of experimental data of a target material and adopting a least square method and a self-service method which are the same as those of an original physical model, and adding prior constraint to use knowledge of the original model; Generating enhanced data, namely generating a large number of virtual samples as a fine tuning training set by using a physical model after recalibration according to the same Latin hypercube sampling strategy in the step 2); and 3) fine tuning of the machine learning model, namely fine tuning training of which the training round is smaller than the limited round trained in the original step 3) is performed on the newly generated enhanced data by taking the original model parameters as initial values, and training is performed by combining an early-stop system.
9. An intelligent regulation and control system applying the integrated intelligent regulation and control method, the system comprises: the data processing module is used for receiving and preprocessing initial small amount of process data; A physical model module, which is internally provided with an original physical model for calibration and executes the step 1); the data enhancement module is used for executing the step 2) to generate virtual data and obtain enhanced data; and 3) executing the step 3) to obtain a trained multi-target prediction model.
10. The intelligent regulation system of claim 9 further comprising: And the process optimization module is used for receiving target performance set by a user, calling the proxy model module, carrying out inverse solution by combining a global optimization algorithm, and outputting recommended optimal process parameters.

Description

Forging heat coordinated material organization and performance integrated intelligent regulation and control method and system Technical Field The invention belongs to the technical field of material processing and intelligent manufacturing, and particularly relates to a steel forging and heat treatment collaborative process optimization and intelligent control method and system combining a physical mechanism model and a data driving model. The invention is particularly suitable for the high-end manufacturing fields of aerospace, automobile industry, precision dies and the like, optimizes the technological parameters and predicts the quality of high-performance carburized steel parts, and can realize high-precision collaborative prediction and fine regulation and control of material organization and performance under the condition of industrial small sample data. Background Forging and heat treatment are key process links for determining final service performance of high-end metal components (particularly carburized steel parts used in the fields of aerospace, precision dies, automobile transmission systems and the like). Vacuum carburization is used as a new generation of advanced heat treatment process, and carbon atoms are permeated into the surface of steel materials in a low-pressure environment to form a high-carbon hardening layer, so that the surface hardness, wear resistance and fatigue strength of parts can be remarkably improved. The traditional process parameter design is highly dependent on expert experience and a large number of trial-and-error experiments, so that the research and development period is long, the cost is high, and the uniformity and consistency of the organization and the performance of the final product in different batches and different positions are difficult to ensure. With the continuous increase of the requirements of high-end manufacture on the quality of products, the establishment of an intelligent model capable of accurately predicting the complex mapping relationship between process parameters, microstructure and macroscopic performance is urgently needed. At present, the carburization process prediction and optimization mainly adopts the following two types of technical paths, but the two types of technical paths have obvious defects: the method simulates a temperature field, a carbon concentration field and a phase change process in the heat treatment process by means of finite element analysis and the like. The theoretical basis comprises: carbon diffusion follows Fick's second law, and under semi-infinite medium boundary conditions, the carbon concentration profile can be expressed as: 。 where Cs is the surface carbon content, D is the diffusion coefficient, t is the diffusion time, and x is the depth from the surface. The diffusion coefficient D follows the Arrhenius temperature dependence: 。 However, this method has the following limitations: Vacuum carburization adopts a pulse carburetion-diffusion circulation process, and the constant boundary condition is assumed to be invalid; The diffusion coefficient is influenced by the carbon concentration and the complexity of alloy elements, and a large error is introduced in the simplification treatment; the model does not consider the influence of martensitic transformation in the quenching process on the final performance; under the industrial condition, the prediction error can reach 20% -40%, and the requirement of fine control is difficult to meet. (II) machine learning method based on data driving The method directly establishes the mapping relation between the technological parameters and the performances by using a machine learning algorithm, but has the following core defects: The data demand is large, hundreds to thousands of groups of training samples are usually needed, and a single carburization experiment takes 4 to 8 hours, and the cost is as high as thousands to tens of thousands of yuan; The small sample overfitting is serious, when the sample size is insufficient (n < 20), the model is easy to memorize noise instead of real rules; The physical interpretability is lost, namely, a black box model cannot explain a prediction mechanism and is difficult to trust by engineering personnel; physical consistency is not guaranteed, and prediction results which violate physical rules can be generated. The core contradiction faced by the prior art is that, on one hand, the requirements of high-end manufacturing on the process prediction precision and reliability are increasingly improved, and on the other hand, the current situation of small sample data (usually only 5-15 groups of experimental data), a complex multi-physical field coupling mechanism and the requirements on physical interpretability which are commonly existed in industrial practice cannot meet the requirements of the prior art scheme. Therefore, there is an urgent need to develop an intelligent regulation method that can integrate physical mec